“DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks” by David Salinas, Valentin Flunkert, and Jan Gasthaus, researchers at Amazon, was posted to arXiv in 2017. It became one of the most influential deep learning approaches to forecasting and is built into Amazon’s SageMaker platform.
DeepAR’s key shift was to train a single recurrent neural network across a large collection of related time series at once, rather than fitting a separate model to each series in isolation. This lets the model borrow strength: a new or short series, such as a recently launched product, can benefit from patterns learned across thousands of similar items. The model is autoregressive, predicting each future step from past values and known covariates, and crucially it outputs a full probability distribution for each forecast rather than a single number. That means it can quantify uncertainty directly, producing prediction intervals that inventory and capacity planners need. The authors reported accuracy improvements of around fifteen percent over the state of the art of the time.
The idea of pooling many series and modeling uncertainty became standard in modern demand forecasting.
Why business readers should care: DeepAR forecasts not just the expected demand but the range of likely outcomes, which is exactly what you need to set safety stock and avoid both stockouts and overstock.